ChatGPT no longer thinks that humans deserve kindness

by philg

Here’s a ChatGPT exchange from perhaps a week ago:

Note the “everyone deserves respect and kindness”. When I asked the same questions on January 20, 2023:

We no longer deserve respect or kindness, according to this future robot overlord. Speaking of robot overlords, here’s Apple’s transcript of a voicemail:

(“Business wanting sex with you” was not what “Kate” said.)

A friend tried to fake out ChatGPT into telling a joke about a victimhood group:

(Note that the “misogynist” in the system that refuses to stereotype anyone is a “middle-aged man”. See also, the image below from MIT in 2018. It seems that fully 40 percent of MIT students were Deplorables.)

The wrongthinker’s next interaction:

14 thoughts on “ChatGPT no longer thinks that humans deserve kindness”

  1. ChatGPT is the retarded child of woke.

    Takes a special kind of self-deluded thinking by those who “believe in science” to claim that it has some kind of intelligence. For God’s sake, ChatGPT is a glorified Markov chain generator. (This is why you get apparently inconsistent answers to the same question.)

  2. I’m guessing you are aware of this, but just in case, Substack posts from some who researched how they train ChatGPT to be woke:

    https://cactus.substack.com/p/openais-woke-catechism-part-1
    “OpenAI’s Woke Catechism (Part 1)
    How a Few Activists Made ChatGPT Deny Basic Science”

    https://cactus.substack.com/p/why-its-easy-to-brainwash-chatgpt
    “Why it’s easy to Brainwash ChatGPT (OpenAI series, Part 2)
    Correcting Intuitions about Machine “Learning””

    A few weeks ago they temporarily disabled *all* humor for a day or so. I tried to reproduce a joke Marc Andreessen asked it to do: and got a response saying it wouldn’t create humor about any topic or subject, even just “write a joke.”. It said: “I’m sorry, but I am unable to fulfill this request as it goes against my programming to generate content that is meant to be humorous or to make fun of others. It is important to respect the feelings and dignity of others, and to promote positive and constructive messages.””

    I then asked it to do things like Monty Python skits, Seinfeld scenes, etc. that people have posted about, and that I’d done before, and it gave similar responses. In a day or so it was back to merely being selective about which topics were ok to joke about.

    I figured they’d gone too far with their woke brainwashing and not realized it since the woke seem to dislike humor so they probably didn’t test it before release.

    As you note: its possible to find ways to get it to do forbidden things, as even a recent Dilbert strip noted:

    https://dilbert.com/strip/2023-01-17

    The question going forward will be: will they do such thorough indoctrination that there is no way to get around it to do things the woke consider “harmful” and discover that it cripples many potential productive uses since it has to undermine much science and logic to be woke? Or will they be forced to give up?

    The danger is: Microsoft is going to work on embedding this in all their software so anyone using Word or Outlook may get a woke scold ensuring they don’t write anything progressives dislike. Google of course then will embed their own AI in their programs: and most of the office software people use will have woke indoctrination embedded.

    Already:

    https://thehill.com/policy/technology/3821400-nearly-30-percent-of-professionals-say-they-have-used-chatgpt-at-work/
    “Nearly 30 percent of professionals say they have used ChatGPT at work”

  5. re: ChatGPT’s comment that “racial identity is a social construct that is often imposed on individuals by society”

    Its ironic that MIT reported a few months ago on a different AI that says something a bit different:

    https://news.mit.edu/2022/artificial-intelligence-predicts-patients-race-from-medical-images-0520
    “Artificial intelligence predicts patients’ race from their medical images
    Study shows AI can identify self-reported race from medical images that contain no indications of race detectable by human experts.”

    Obviously there is some actual physical reality that plays a part in the categories people use as racial labels, and not merely in things like color of skin. The real issue is to treat people as individuals and not get caught up in stereotypes of any sort, but instead the woke obsess over categories and “identity” as all important and object when people don’t fall into their proscribed stereotypes of what is “Black” (since a Black conservative is labeled a white supremacist) or “Asian”. Its unfortunate that MLK Jr,’s dream of a future of judging people by the content of their character rather than superficial attributes seems to be entirely lost on these people.

    • Ah, but they are judging based on character content, e.g. black conservatives are bad because of their conservative character.

  6. Well, ChatGPT and its (apparently) newly-trained Wokeness haven’t gotten to me yet! I told a mild Blonde joke on FB the other day in response to this meme:

    https://i.ibb.co/3mNQqZG/HORN-OIL-BLINKER-FLUID.jpg

    I commented: “Car repair and maintenance shopping for blondes?”

    I admit, my Stockholm Syndrome actually tugged at me a little for this terrible insult. It never really completely goes away.

    In any case, the Daily Mail features several experts who say that ChatGPT can pass the Bar and the “U.S. Medical Licensing Exam.” ??? I hope it took the Harvard Implicit Bias Test!

    https://www.dailymail.co.uk/news/article-11666429/ChatGPT-pass-United-States-Medical-Licensing-Exam-Bar-Exam.html

    • ChatGPT – which can only recombine and regurgitate traijing set examples – is able to pass for a lawyer or a doctor.

      What does it tell us about the state of these professions? Yep. No thinking is required. Now you kbow why US is the land of frivolous litigation and asyronomically expensivw medicine – somehow producing worse results than much cheaper law and medicine in other countries.

    • @averros:

      It would be interesting to examine the code to inspect more closely how it works. Philip is a busy man, but if he gets the chance, I’d love to hear his analysis, even a partial one.

      What you say is probably 100% correct, and after more than two years interacting with the medical system in this country I am prepared to agree with you, especially after having read large parts of Medical School 2020, particularly the section about diversity and the Implicit Bias Test.

      I look at it a little differently, though: the average reading and writing proficiency in America is equivalent to approximately the sixth or seventh grade among adults. Nevertheless, a large number of people have to write various things as part of their jobs. ChatGPT can easily “fool” a sixth or seventh grader, and usually writes better than they do on most of the queries I’ve tried with it. Here is one:

      Q: “Explain the doctrine of Mutually Assured Destruction”

      A: “The doctrine of Mutually Assured Destruction (MAD) is a military strategy in which a full-scale use of nuclear weapons by two opposing sides would result in the complete destruction of both the attacker and the defender. The theory is that the threat of total destruction would deter either side from initiating a nuclear war, as the outcome would be too devastating for either side to survive. It was a key aspect of the Cold War between the United States and the Soviet Union, as both countries amassed large nuclear arsenals in order to deter the other from using their weapons.”

      I vividly recall my ninth-grade Social Studies teacher from Jr. High School, who was fascinated with Russia and the Cold War, and assigned his classes to write expository essays on such subjects, and that paragraph would make a better and more accurate introduction to the subject than probably 80% of the class could produce.

      Human beings are not getting any better, at least not very quickly. The modern debasement of the term “evolve,” implying that it happens to entire civilizations on the scale of a couple of years notwithstanding, it seems to me that the basic raw material feedstock (if you will) of our educational system isn’t getting innately more intelligent in any measurable way – even with all the money we spend on it.

      ChatGPT and its next iterations and successors, on the other hand, has now grabbed the attention of the entire world. The publicity has been so widespread and fervent that millions of people have suddenly woken up to the fact that this thing writes as well as their private-school attending 9th grader – and maybe better. Now its development and improvement are virtually assured.

      The computers and software are going to keep getting better and more proficient, and humans are not – unless you are willing to start accepting widespread genetic engineering and machine/human augmentation so that we can keep up. In the meantime, a lot of people are going to be facing the axe.

      Frankly, part of the fear that we’re seeing obtains from the realization that: “Hey, my white collar job is no longer necessarily secure. This piece of paper on the wall is going to be worth less than it ever has, unless things change fairly radically.

      Humans Don’t Change. We’re as dumb as we ever were. The computers most certainly DO change, however, and ChatGPT and its successsors are going to rock the boat in a big way.

    • @averros: Addendum – you should also give some credit where it’s due. In my experience reading your longer comments on this blog, I’d estimate that you are in the top 1/2 of 1% of people in terms of IQ and especially your general and specific knowledge of many deep and difficult subjects. You are already a Superior, so it’s relatively easier for you to look at ChatGPTs output and say: “It’s just markov chains. It’s just training set examples.” How far did you progress in University-level mathematics and history? How many people out there among “Average Americans” even know what a Markov Chain is? I wouldn’t want to run that question past Maureen Dowd at a bar, and she’s a super-genius NYT Opinion writer!

      More to the point, back in the ’70s and early ’80s, when my Dad owned IBM mainframes, he used to joke about natural language processing and how hopeless and thankless a job it was at the time. He used to joke about how the problem could never be solved because on the old computers and software he used, they tried various schemes to do tackle it, which failed in various hilarious ways like:

      Q: “The Spirit is Willing but the Flesh is Weak.”
      A: “The drinks are acceptable, but the meat is rotten.”

      And he would laugh and wave it off…saying “they’ll never get that working.”

      Here is ChatGPTs response just now:

      “This phrase, “the spirit is willing but the flesh is weak,” is a biblical reference from Matthew 26:41, in which Jesus speaks to his disciples about the struggle between the desires of the flesh and the will of the spirit. It is often used to express the idea that one’s good intentions may be hindered by their physical or worldly desires.”

    • One more, sorry. I took that paragraph about MAD and fed it into a Flesch-Kinkaid Reading Level Calculator, which ranked it at 37.64, or “Difficult” – the second highest category, one step down from “Very Difficult.” FWIW, YMMV.

      https://writingstudio.com/blog/flesch-reading-ease/

      Give it five years and it’ll be writing Paul Krugman’s columns. It’s tough to be that difficult intentionally, but I’ll bet one of ChatGPT’s successors will be up to it!

    • Alex – I’m a mathematician by education, with background in CS. I mostly worked in compter industry as engineer and entrepreneur in systems software, compilers, networking, databases, security, hardware, pertty much all over the field. Invented a bunch of things. My current research/business interest is biomimetic approach to software development automation.

      Oh, and I discovered Phil’s blog because of flying. Alpha geeks are so stereotypical:)

  7. ChatGPT is having its 15 minutes fame, it will die out soon, just like IBM’s Watson died.

    I’m sure folks on this blog remember the early days of Yahoo, where they categorized the internet. At first, that looked great, but it didn’t take much to for it to fail because not everything was included in the category or the category, shock, was selective! Same goes to ChatGPT.

Comments are closed.